阅读说明：本技术 基于独立向量分析的高密度表面肌电信号消噪方法 (High-density surface electromyographic signal denoising method based on independent vector analysis ) 是由 陈勋 王堃 吴乐 张旭 陈香 于 2019-10-15 设计创作，主要内容包括：本发明公开了一种基于独立向量分析的高密度表面肌电信号消噪方法,其步骤包括：1、首先将高密度表面肌电信号通过延时构造多个数据集；2、利用独立向量分析法进行联合盲源分离,得到每个数据集相对应的源信号矩阵和解混矩阵；3、选择高密度表面肌电信号相对应的源信号矩阵和解混矩阵；4、计算源信号矩阵中各个独立源成分肌肉收缩段和静息段的强度比；5、设定消噪阈值,将强度比低于消噪阈值的独立源成分置零；6、盲源分离逆变换得到消除噪声后的干净高密度表面肌电信号。本发明基于独立向量分析,通过设定消噪阈值,在去除工频干扰和高斯白噪声的同时,保证了肌电信息在处理过程中尽可能不丢失。(The invention discloses a high-density surface electromyographic signal denoising method based on independent vector analysis, which comprises the following steps: 1. firstly, constructing a plurality of data sets by delaying high-density surface electromyographic signals; 2. performing combined blind source separation by using an independent vector analysis method to obtain a source signal matrix and a demixing matrix corresponding to each data set; 3. selecting a source signal matrix and a demixing matrix corresponding to the high-density surface myoelectric signals; 4. calculating the intensity ratio of each independent source component muscle contraction section to each independent source component muscle rest section in the source signal matrix; 5. setting a noise elimination threshold value, and setting the independent source components with the intensity ratio lower than the noise elimination threshold value to be zero; 6. and (4) carrying out blind source separation inverse transformation to obtain a clean high-density surface myoelectric signal after noise elimination. The method is based on independent vector analysis, and guarantees that the electromyographic information is not lost as much as possible in the processing process while removing power frequency interference and Gaussian white noise by setting a noise elimination threshold.)

1. A high-density surface electromyographic signal denoising method based on independent vector analysis is characterized by comprising the following steps:

the method comprises the following steps: the myoelectricity measuring device collects and records a high-density surface myoelectricity signal matrix of the N channels at the time t, and the matrix is recorded as follows: x (t) ═ x₁(t),x₂(t),...,x_n(t),...,x_N(t)]^T，x_n(t) is the electromyographic signal of the nth channel at the time t, and if t belongs to [0, t ∈₁]Then the high-density surface electromyographic signal matrix X (t) is a rest section without electromyographic activity; if t e (t)₁,t₂]Then the high density surface myoelectric signal matrix X (T) is the muscle contraction section containing myoelectric activity, T is the transposition of the matrix, T₁As the moment when the muscle begins to contract, t₂N is more than or equal to 1 and less than or equal to N, t is more than or equal to 0 and less than or equal to t at the acquisition ending moment₂；

Step two: constructing K data sets X (X) by delaying K-1 points of the high-density surface myoelectric signal matrix X (t)^[1],X^[2],…,X^[k]...,X^[K]]Wherein, in the step (A),

step three: performing combined blind source separation on the K data sets X by using an independent vector analysis method to obtain a t-time source signal matrix S (t) and an inverse mixing matrix W of the high-density surface myoelectric signal matrix X (t)^*Wherein s (t) ═ s₁(t),s₂(t),...,s_n(t),...,s_N(t)]^T，s_n(t) represents the nth source signal at time t and has: s (t) ═ W^*X(t)；

Step four: calculating the nth independent source signal s in the source signal matrix S (t) by using the formula (1)_n(t) intensity ratio IR of muscular contraction and resting segment_n：

In the formula (1), c_n(i) Representing the nth source signal s at time t_n(t) sampling point i of muscle contraction segment; q. q.s_n(i) Representing the nth source signal s at time t_n(t) sample point i of the rest segment, m₁＝t₁×f_s，m₂＝(t₂-t₁)×f_s，f_sIs the sampling frequency of the signal;

step five: setting a noise-canceling threshold by comparing the intensity ratio IR_nN-th source signal s less than noise-cancelling threshold_n(t) judging the signal as Gaussian white noise or power frequency interference component, and converting the nth source signal s_n(t) setting to zero; will intensity ratio IR_nNth source signal s greater than noise cancellation threshold_n(t) determining the myoelectric information component and retaining the nth source signal s_n(t)；

Step six: repeating the fourth step and the fifth step to obtain a t-time source signal matrix after the noise component is set to zero

in the formula (3), (W)^*)^-1Is an inverse mixing matrix W^*The inverse matrix of (c).

2. The method for denoising the high-density surface myoelectric signal according to claim 1, wherein the third step is performed as follows:

step 3.1, performing combined blind source separation on the K data sets X by utilizing a multivariate Gaussian distributed independent vector analysis method to obtain K initial unmixing matrixes

step 3.2, utilizing the initial unmixing matrix

step 3.3, obtaining the kth data set X by using the formula (3)^[k]To be estimated source signal matrix

In formula (3), K is 1, 2., K; when k is equal to 1, the first step is carried out,

Technical Field

The invention belongs to the technical field of electromyographic signal processing, and particularly relates to a method for automatically identifying power frequency interference and white Gaussian noise from a high-density surface electromyographic signal based on independent vector analysis, which is mainly applied to preprocessing of the electromyographic signal.

Background

Surface Electromyography (sEMG) is a weak electrophysiological signal recorded by electrodes attached to the skin Surface of a human body that can reflect information about the activity of the neuromuscular system. sEMG signals have been widely used in the fields of neurophysiology, clinical medicine, motor medicine, and rehabilitation medicine. The High Density surface electromyography (HD-sEMG) technique is a spatio-temporal variant of the single channel surface electromyography technique, having a large number of closely spaced electrode channels, providing information on the spatio-temporal distribution of electromyographic intensities over a wide muscle area. Such information can improve muscle force estimation, myoelectricity decomposition, myoelectricity prosthesis control, and the like. However, the electromyographic signals, which are weak electrophysiological signals, are often interfered by noises such as power frequency interference, white gaussian noise, motion artifacts, and the like, thereby affecting the accuracy of subsequent electromyographic analysis.

In the past decades, researchers have proposed various methods for eliminating noise in electromyographic signals. A commonly used method is to eliminate power frequency interference of 50Hz by using a power frequency wave trap, but simultaneously, effective information of the electromyographic signals is greatly influenced, because the frequency range of the electromyographic signals is mainly distributed between 10Hz and 500Hz, irreparable loss is caused to the electromyographic signals during filtering. The wavelet transform has good time-frequency analysis capability, has incomparable superiority with the traditional denoising method, and is widely applied to signal denoising. However, the effect of wavelet de-noising is directly related to the selection of wavelet basis functions, and once the wavelet basis functions are selected, the wavelet basis functions cannot be changed in the process of signal analysis, so that the wavelet transform is poor in self-adaptation. Empirical mode decomposition decomposes a signal into superposition of a plurality of inherent mode functions, each inherent mode function has characteristics of different scales, and the method has good adaptivity, but has the defect of mode mixing. Ensemble empirical mode decomposition adds white noise to the electromyographic signal prior to signal decomposition, which, while overcoming the disadvantages of mode mixing, also contaminates the signal with residual noise.

In addition, blind source separation methods are also commonly used for noise cancellation of biomedical signals, the purpose of which is to separate different signal sources from the observed signal and to obtain a mixing matrix. Among the most common blind source separation methods are Independent Component Analysis (ICA) and Canonical Correlation Analysis (CCA). The principle of applying the blind source separation method to noise elimination in the electromyographic signals is that a series of independent source components are extracted from multi-channel observation signals, then source components representing noise in the source components are set to zero, and finally the electromyographic signals after noise elimination are obtained through reverse blind source separation. The CCA decomposes the electromyographic signal into typical variables that are not correlated with each other but have the greatest autocorrelation using second-order statistics. Because the frequency band of the Gaussian white noise is wide and has a relatively low autocorrelation coefficient compared with the electromyographic signals, the Gaussian white noise component and the electromyographic component can be well separated by utilizing the second-order statistical characteristic. However, only the typical variables are required to be uncorrelated with each other and the power frequency interference components cannot be completely separated, so that the CCA method has the problem that the myoelectric signal components are also mixed and have power frequency interference. ICA decomposes the electromyographic signals into mutually independent components by using high-order statistics, and can better separate the power frequency interference component from the electromyographic signal component, but the separated power frequency interference component still contains more electromyographic information. In addition, because ICA cannot separate gaussian components from each other, directly zeroing the separated gaussian components cannot achieve the goal of eliminating white gaussian noise.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention provides the high-density surface electromyographic signal denoising method based on independent vector analysis, so that the influence of power frequency interference and Gaussian white noise on the electromyographic signal can be completely removed, and meanwhile, the information of the electromyographic component is kept as far as possible without being lost, thereby improving the analysis accuracy of the electromyographic signal and providing a new idea for electromyographic denoising.

In order to solve the technical problem, the invention adopts the following technical scheme:

the invention relates to a high-density surface electromyographic signal denoising method based on independent vector analysis, which is characterized by comprising the following steps of:

the method comprises the following steps: the myoelectricity measuring device collects and records a high-density surface myoelectricity signal matrix of the N channels at the time t, and the matrix is recorded as follows: x (t) ═ x₁(t),x₂(t),...,x_n(t),...,x_N(t)]^T，x_n(t) is the electromyographic signal of the nth channel at the time t, and if t belongs to [0, t ∈₁]Then the high-density surface electromyographic signal matrix X (t) is a rest section without electromyographic activity; if t e (t)₁,t₂]Then the high density surface myoelectric signal matrix X (T) is the muscle contraction section containing myoelectric activity, T is the transposition of the matrix, T₁As the moment when the muscle begins to contract, t₂N is more than or equal to 1 and less than or equal to N, t is more than or equal to 0 and less than or equal to t at the acquisition ending moment₂；

Step two: constructing K data sets X (X) by delaying K-1 points of the high-density surface myoelectric signal matrix X (t)^[1],X^[2],…,X^[k]…,X^[K]]Wherein, in the step (A),

a representation of the k-th data set,

represents the kth data set X^[k]The electromyographic signal of the nth channel, K is 1,2,. and K;

step three: performing combined blind source separation on the K data sets X by using an independent vector analysis method to obtain a t-time source signal matrix S (t) and an inverse mixing matrix W of the high-density surface myoelectric signal matrix X (t)^*Wherein s (t) ═ s₁(t),s₂(t),...,s_n(t),...,s_N(t)]^T，s_n(t) represents the nth source signal at time t and has: s (t) ═ W^*X(t)；

Step four: calculating the nth independent source signal s in the source signal matrix S (t) by using the formula (1)_n(t) intensity ratio IR of muscular contraction and resting segment_n：

In the formula (1), c_n(i) Representing the nth source signal s at time t_n(t) sampling point i of muscle contraction segment; q. q.s_n(i) Representing the nth source signal s at time t_n(t) sample point i of the rest segment, m₁＝t₁×f_s，m₂＝(t₂-t₁)×f_s，f_sIs the sampling frequency of the signal;

step five: setting a noise-canceling threshold by comparing the intensity ratio IR_nN-th source signal s less than noise-cancelling threshold_n(t) judging the signal as Gaussian white noise or power frequency interference component, and converting the nth source signal s_n(t) setting to zero; will intensity ratio IR_nNth source signal s greater than noise cancellation threshold_n(t) determining the myoelectric information component and retaining the nth source signal s_n(t)；

Step six: repeating the fourth step and the fifth step to obtain a t-time source signal matrix after the noise component is set to zero

Obtaining the denoised electromyographic signal by using the blind source separation inverse transformation shown in the formula (2)

The denoised electromyographic signal of the nth channel at the time t;

in the formula (3), (W)^*)^-1Is an inverse mixing matrix W^*The inverse matrix of (c).

The method for eliminating the noise of the high-density surface electromyographic signals is also characterized in that the third step is carried out according to the following process:

step 3.1, performing combined blind source separation on the K data sets X by utilizing a multivariate Gaussian distributed independent vector analysis method to obtain K initial unmixing matrixes

Is the kth data set X obtained by independent vector analysis method of multivariate Gaussian distribution^[k]The initial unmixing matrix of (a);

step 3.2, utilizing the initial unmixing matrixInitializing an independent vector analysis method of multivariate Laplace distribution, and performing combined blind source separation on the K data sets X to obtain K unmixing matrixes W [ W [ ] [^[1],W^[2],...,W^[k]...,W^[K]]，W^[k]Is the kth data set X obtained by independent vector analysis of multivariate Laplace distribution^[k]The unmixing matrix of (a);

step 3.3, obtaining the kth data set X by using the formula (3)^[k]To be estimated source signal matrix

In formula (3), K is 1, 2., K; when k is equal to 1, the first step is carried out,the source signal at the t moment obtained by the high-density surface myoelectric signal matrix X (t) through an independent vector analysis methodNumber matrix, denoted as s (t) ═ s₁(t),s₂(t),...,s_n(t),...,s_N(t)]^TWherein s is_n(t) the nth source signal at time t; w^[1]Is an inverse mixing matrix of the high-density surface myoelectric signal matrix X (t) and is marked as W^*；X^[1]Is the high density surface myoelectric signal matrix X (t).

Compared with CCA and ICA methods, the method can remove the influence of power frequency interference and white Gaussian noise on myoelectricity, and can retain myoelectricity information to the maximum extent, and the specific beneficial effects are as follows:

1. in the second step and the third step of the invention, a plurality of data sets are obtained through time delay, and electromyographic signals of the data sets are simultaneously subjected to combined blind source separation by utilizing independent vector analysis, compared with ICA and CCA, the method can provide more source signal related information for blind signal separation, and promote a power frequency interference source, a Gaussian white noise source and an electromyographic source to be more accurately separated into different independent components, so that in the subsequent fourth step and the fifth step, when the noise component is automatically removed through a threshold value, the component loss of the electromyographic signals is less.

2. In the third step of the invention, when the independent vector analysis is used for carrying out the joint blind source separation on the signals of multiple data sets, the separated source signals are mutually independent in the same data set, and the corresponding source signals have the maximum correlation in different data sets. Compared with the ICA method, the method considers the second-order statistical characteristic of the Gaussian white noise, so that the separation of the Gaussian white noise is more effective, meanwhile, the separated power frequency interference component also contains less electromyographic information, and the loss of the electromyographic signal information in the denoising process is reduced. Compared with the CCA method, the power frequency interference source signal obtained by separation is statistically independent from the myoelectric signal source, so that the power frequency interference source is separated more thoroughly, and the problem of incomplete removal of the power frequency interference in the CCA method is solved.

3. In the third step, compared with the existing blind source separation method based on ICA and CCA, the independent vector analysis method utilizes the hypothesis that the source component vector is in multivariate Gaussian distribution and multivariate Laplacian distribution, unifies the second-order statistic and the high-order statistic in the mathematical model of the combined blind source separation, fully combines the advantages of the ICA and CCA methods, fully considers the respective statistical characteristics of the electromyographic signal and the noise signal, can well separate a Gaussian white noise source component, a power frequency interference source component and an electromyographic signal source component, and better solves the problem of noise elimination in the electromyographic signal.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2a is a schematic representation of the first 2 seconds for the first 20 channels to simulate a high density surface myoelectric signal;

FIG. 2b is a schematic of the first 2 seconds of the first 20 channels of a high density surface myoelectric signal contaminated with two types of noise;

FIG. 3 is a schematic representation of the first 2 seconds of the first 20 channels of a source signal matrix obtained by the method of the present invention;

FIG. 4 is a graph illustrating the ratio of the intensity of the source signal obtained by the method of the present invention;

FIG. 5 is a schematic representation of the first 2 seconds of a source signal after noise cancellation by the method of the present invention;

FIG. 6 is a schematic representation of the first 2 seconds for the first 20 channels of a high density surface electromyographic signal after noise cancellation by the method of the invention

FIG. 7 is a comparison graph of the denoising performance RRMSE of the CCA and ICA methods according to the present invention;

FIG. 8a is a comparison graph of the noise reduction performance signal-to-noise ratio of simulated myoelectricity processed by the method of the present invention and CCA and ICA methods

FIG. 8b is a waveform comparison diagram of a signal after denoising in the method of the present invention and CCA and ICA methods for processing simulated myoelectricity;

FIG. 9a is a schematic of the first 2 seconds of the first 20 channels of a 48 channel true high density surface myoelectric signal;

FIG. 9b is a schematic representation of the first 2 seconds for the first 20 channels of the 48-channel surface electromyographic signals denoised by the method of the present invention;

FIG. 10a is a comparison graph of the noise reduction performance signal-to-noise ratio of the method of the present invention and CCA and ICA methods for processing real myoelectricity;

FIG. 10b is a waveform comparison diagram of a signal after denoising in the method of the present invention and CCA and ICA methods for processing real myoelectricity;

FIG. 11a is a spectrum diagram of a signal before denoising;

FIG. 11b is a graph showing a comparison of spectral regions before and after noise removal by the ICA method;

fig. 11c is a comparison graph of the spectrum before and after noise removal by the CCA method;

FIG. 11d is a graph showing a comparison of spectral regions before and after noise removal by the IVA method.

Detailed Description

In this embodiment, as shown in fig. 1, a method for denoising a high-density surface myoelectric signal based on independent vector analysis includes: firstly, delaying high-density surface myoelectric signals to obtain a plurality of data sets, and then performing combined blind source separation on the data sets by using an independent vector analysis method; then obtaining a source signal matrix and an inverse mixing matrix of the high-density surface myoelectric signals; then calculating the intensity ratio of each independent source component muscle contraction section to each independent source component muscle rest section in the source signal matrix, setting the source signals lower than the noise elimination threshold value to zero, and reserving the source signals higher than the noise elimination threshold value; and finally, carrying out blind source separation and inverse transformation to obtain the high-density surface myoelectric signal after noise elimination.

The following describes specific embodiments by taking an example of a simulated electromyographic signal and a real electromyographic signal, respectively, with reference to the accompanying drawings.

1 analog electromyographic signal

In this section, two examples will be described, the first example being to describe a specific embodiment of the present invention, and the second example being to compare the present invention with the ica (fastica) and CCA processing methods.

(1) Example one

The method comprises the following steps: the myoelectricity measuring device collects and records a high-density surface myoelectricity signal matrix of the N channels at the time t, and the matrix is recorded as follows: x (t) ═ x₁(t),x₂(t),...,x_n(t),...,x_N(t)]^T，x_n(t) is the electromyographic signal of the nth channel at the time t, and if t belongs to [0, t ∈₁]Then the high-density surface electromyographic signal matrix X (t) is a rest section without electromyographic activity; if t e (t)₁,t₂]Then high density surface myoelectric signal momentThe matrix X (T) is the muscle contraction segment containing myoelectric activity, T is the transposition of the matrix, T₁As the moment when the muscle begins to contract, t₂N is more than or equal to 1 and less than or equal to N, t is more than or equal to 0 and less than or equal to t at the acquisition ending moment₂；

In this embodiment, a myoelectric measurement device collects and records a high-density surface myoelectric signal matrix with 64 channels at time N, which is recorded as: x (t) ═ x₁(t),x₂(t),...,x_n(t),...,x₆₄(t)]^T，x_n(t) is the electromyographic signal of the nth channel at the time t, and if t belongs to [0,0.5 ]]Then the high-density surface electromyographic signal matrix X (t) is a rest section without electromyographic activity; if t is an element of (0.5, 5.5)]Then the high density surface myoelectric signal matrix X (T) is the muscle contraction section containing myoelectric activity, T is the transposition of the matrix, T₁0.5 is the time at which the muscle begins to contract, t₂N is more than or equal to 1 and less than or equal to 64, and t is more than or equal to 0 and less than or equal to 5.5 at the moment of finishing acquisition; x (t) X as shown in fig. 2b_EMG(t)+X_WGN(t)+X_PLI(t) wherein X_EMG(t)＝[x_EMG1(t),x_EMG2(t),...,x_EMG64(t)]^TThe simulated high density surface myoelectric signal matrix representing 64 channels, as clearly shown in FIG. 2a, t e 0,0.5]The signal amplitude is zero, is a resting segment without myoelectric activity, 0.5 second is the moment when the muscle begins to contract, the signal amplitude is gradually increased, and t belongs to (0.5, 5.5)]Being muscular contraction segments containing myoelectric activity, X_WGN(t)＝[x_WGN1(t),x_WGN2(t),...,x_WGN64(t)]^TSimulated white Gaussian noise signal, X, representing 64 channels_PLI(t)＝[x_PLI1(t),x_PLI2(t),...,x_PLIch(t)]^TCh is more than or equal to 15 and less than or equal to 20, and the simulated power frequency interference signal of the ch channel is represented; the specific simulation process of X (t) is as follows: first add 64 channels of Gaussian white noise X_WGN(t) such that the signal-to-noise ratio of 64 channels of signal x (t) follows a normal distribution of N (20,3) after adding white gaussian noise; then adding power frequency interference X_PLIAnd (t), randomly selecting ch channels from 64 channels to add power frequency interference, so that the channels added with the power frequency interference in the signal X (t) obey normal distribution with the signal-to-noise ratio of N (7, 3).

Step two: surface of high densityElectromyographic signal matrix X (t) constructs K data sets X-X through delaying K-1 points^[1],X^[2],…,X^[k]…,X^[K]]Wherein, in the step (A),

a representation of the k-th data set,represents the kth data set X^[k]The electromyographic signal of the nth channel, K is 1,2,. and K;

in this embodiment, K is 21, and a high-density surface myoelectric signal matrix X (t) is delayed by K-1 points to construct 21 data sets X ═ X^[1],X^[2],…,X^[k]…,X^[21]]Wherein, in the step (A),

a representation of the k-th data set,represents the kth data set X^[k]Electromyographic signals of the nth channel, k ═ 1, 2., 21;

step three: performing combined blind source separation on the K data sets X by using an independent vector analysis method to obtain a t-time source signal matrix S (t) and an inverse mixing matrix W of a high-density surface myoelectric signal matrix X (t)^*Wherein s (t) ═ s₁(t),s₂(t),...,s_n(t),...,s_N(t)]^T，s_n(t) represents the nth source signal at time t and has: s (t) ═ W^*X (t); the method comprises the following specific steps:

step 3.1, performing combined blind source separation on the K data sets X by utilizing a multivariate Gaussian distributed independent vector analysis method to obtain K initial unmixing matrixes

For independent vector analysis by multivariate Gaussian distributionThe k-th data set X^[k]The initial unmixing matrix of (a);

step 3.2, utilizing the initial unmixing matrix

Initializing an independent vector analysis method of multivariate Laplace distribution, and performing combined blind source separation on K data sets X to obtain K unmixing matrixes W ═ W^[1],W^[2],…,W^[k]…,W^[K]]，W^[k]Is the kth data set X obtained by independent vector analysis of multivariate Laplace distribution^[k]The unmixing matrix of (a);

step 3.3, obtaining the kth data set X by using the formula (3)^[k]To be estimated source signal matrix

In formula (3), K is 1, 2., K; when k is equal to 1, the first step is carried out,a source signal matrix at the time t obtained by an independent vector analysis method for the high-density surface myoelectric signal matrix X (t) is recorded as S (t) ═ s₁(t),s₂(t),...,s_n(t),...,s_N(t)]^TWherein s is_n(t) the nth source signal at time t; w^[1]An inverse mixing matrix of the high-density surface myoelectric signal matrix X (t), denoted as W^*；X^[1]Is a high density surface myoelectric signal matrix X (t).

In this embodiment, 21 data sets X are subjected to joint blind source separation by using an independent vector analysis method to obtain a source signal matrix s (t) at time t of the high-density surface myoelectric signal matrix X (t), as shown in fig. 3, because the independent vector analysis method based on multivariate gaussian distribution considers the second-order correlation of corresponding source signals among the data sets, the separated source signals are automatically arranged from large to small according to autocorrelation coefficients, it can be seen from fig. 3 that a power frequency interference component with the largest autocorrelation coefficient is a first separated source signal, an independent source signal distributed behind is a gaussian white noise component with the lowest autocorrelation coefficient, and an independent source signal between the two is a myoelectric signal component. Since the independent vector analysis method of the multivariate Laplace distribution utilizes high-order statistics, and the separated source signals are mutually independent in the same data set, the power frequency interference component and the myoelectric signal component can be completely separated, and as can be seen from FIG. 3, only the first source signal is the power frequency interference component, and other myoelectric signal components do not contain the power frequency interference component.

Step four: calculating the nth independent source signal s in the source signal matrix S (t) by using the formula (1)_n(t) intensity ratio IR of muscular contraction and resting segment_n：

In the formula (1), c_n(i) Representing the nth source signal s at time t_n(t) sampling point i of muscle contraction segment; q. q.s_n(i) Representing the nth source signal s at time t_n(t) sample point i of the rest segment, m₁＝t₁×f_s，m₂＝(t₂-t₁)×f_s，f_sIs the sampling frequency of the signal;

in this embodiment, the intensity ratio of the muscle contraction segment and the resting segment of the 64 independent source signals in the source signal matrix s (t) is calculated by using the formula (1), as shown in fig. 4, wherein the IR values of the independent source signals are arranged in a descending order. The larger the IR value is, the more myoelectric activities contained in the muscle contraction section of the source signal are, as shown in fig. 3 from the 2 nd source signal to the 11 th source signal, it can be obviously seen that the signal rest section contains a small amount of noise, and the muscle contraction section contains a large amount of myoelectric activities; the smaller the IR value is, the more noise components are contained in the source signal, because the signal amplitudes of Gaussian white noise and power frequency interference in the muscle contraction section and the muscle contraction section are stable and have no obvious change, as shown in FIG. 3, the amplitudes of the two kinds of noise in the muscle contraction section and the muscle contraction section are stable, and the calculated IR value is about 1. As shown in fig. 3 and 4, it can be seen that the IR value of the noise component is very low, so that the noise elimination threshold is set to remove the noise component and simultaneously retain most of the myoelectric component, thereby ensuring that the myoelectric information is not lost as much as possible.

Step five: setting a noise-canceling threshold by comparing the intensity ratio IR_nN-th source signal s less than noise-cancelling threshold_n(t) judging the signal as Gaussian white noise or power frequency interference component, and converting the nth source signal s_n(t) setting to zero; will intensity ratio IR_nNth source signal s greater than noise cancellation threshold_n(t) determining the myoelectric information component and retaining the nth source signal s_n(t)；

In this embodiment, a denoising threshold value is set to 1.1, and a source signal matrix after denoising is obtained as shown in fig. 5, it can be seen that power frequency interference and most of gaussian white noise are eliminated, the retained 1 st to 10 th source signals obviously contain rich myoelectric activity, while the 11 th to 17 th source signals contain more myoelectric information components although containing white noise, and when the source signals are removed, the signal-to-noise ratio of the myoelectric signals is lowered, so that the retention of the source signals can ensure that the myoelectric information is not lost as much as possible;

step six: repeating the fourth step and the fifth step to obtain a t-time source signal matrix after the noise component is set to zero

Obtaining the denoised electromyographic signal by using the blind source separation inverse transformation shown in the formula (2)

De-noised electromyographic signals of the nth channel at the time t

As shown in figure 5 of the drawings,

in the formula (3), (W)^*)^-1Is an inverse mixing matrix W^*The inverse matrix of (c).

In this embodiment, a denoised electromyographic signal is obtained

As shown in fig. 6, it is clear from fig. 6 that the white gaussian noise of each channel is almost completely eliminated, and comparing with fig. 2b, the power frequency interference of the 9 th, 12 th, 13 th, 15 th, 17 th and 19 th channels is also completely eliminated; therefore, the method can effectively remove the white Gaussian noise, simultaneously has very thorough separation of power frequency interference, and ensures that the electromyographic information is not lost as much as possible while removing the noise.

(2) Example two

To quantitatively assess the effect of the present invention, the inventive method (IVA) was compared to two blind source separation algorithms, ica (fast ica) and CCA, for this purpose. Two performance indexes, namely a Relative Root Mean Square Error (RRMSE) and a signal-to-noise ratio (SNR), are selected as evaluation indexes. The relative root mean square error is defined as follows:

the signal-to-noise ratio is defined as:

SNR_nrepresenting the signal-to-noise ratio, x, of the nth channel_n(i) Denotes the ith sample point, q, of the nth channel in X (t)_n(i) I-th sample point representing the rest segment of the nth channel in x (t), N ═ 1,2₁＝t₁*f_s，m＝t₂*f_s，f_sIs the sampling frequency of the signal.

The smaller the RRMSE value is, the better the RRMSE value is, the smaller the difference between the de-noised signal and the original clean electromyographic signal is; the larger the SNR value, the better, the larger the power representing the noise, and the cleaner the noise removal.

According to the steps in example 1, the simulation experiment is repeated for 30 times, the RRMSE of the 30 simulation experiments is averaged, and RRMSE comparison graphs of the three methods are drawn, as shown in FIG. 7, it can be seen that ICA and CCA have equivalent noise elimination capability but are obviously weaker than IVA method, and the RRMSE value of IVA method is in a slow descending trend along with the increase of the number of data sets, that is, the noise elimination capability of IVA method is slowly enhanced along with the increase of the number of data sets.

Fig. 8a shows the snr conditions of the channels before and after the noise is eliminated by the three methods in the primary simulation experiment, and it can be seen from the figure that the IVA method can greatly improve the snrs of all channels, while the ICA and CCA methods improve the snr of the channel with low snr and reduce the snr of the channel with high snr. Fig. 8b is a comparison graph of EMG signals before and after removing noise in the 38 th channel, and it can be seen that the IVA algorithm can well remove white gaussian noise and power frequency interference. The CCA can not completely remove power frequency interference, ICA cannot completely separate Gaussian white noise, and the defects of the CCA and the ICA cause the signals after blind source separation inverse transformation to still have obvious power frequency interference and Gaussian white noise interference, so that the signal-to-noise ratio of a high signal-to-noise ratio channel is reduced.

2 true electromyographic signals

The real electromyographic data is used as an experimental object, processed by an IVA algorithm, compared with an ICA algorithm and a CCA algorithm, and the denoising effect of the method is judged. Fig. 9a shows a 48-channel real electromyographic signal, wherein the first 0.5 second is a resting segment and the last 5 seconds is a muscle activity segment. It can be seen from fig. 9a that the signal is severely polluted by white gaussian noise and power frequency interference, and it can be observed that the interference degree of each channel is different, so it is necessary to eliminate power frequency interference and increase the signal-to-noise ratio of the channel.

According to the steps of the method of the present invention, the real electromyographic signals are processed, and the reconstructed electromyographic signals after denoising are shown in fig. 9 b. By comparing fig. 9a and fig. 9b, it can be found that the invention can remove the power frequency interference completely, and can completely retain the information of the electromyographic signal.

In addition, the above-mentioned real electromyographic signals are processed by using CCA and ICA methods, and a signal-to-noise ratio of the noise-removed front and rear channels is obtained as shown in fig. 10 a. By contrast, it can be found that the CCA method has the same noise removal capability as the ICA method, and the signal-to-noise ratio of some channels is not significantly improved by the ICA method, while the invention can stably improve the signal-to-noise ratio of all channels. Fig. 10b is a comparison graph of signals before and after the surface electromyogram signal of the 48 th channel and the 22 nd channel remove noise in a time domain, and it can be found that three methods can effectively remove power frequency interference. However, the CCA method cannot completely remove the power frequency interference component, and the ICA method introduces a small amount of interference while removing the power frequency interference, so that the signal-to-noise ratio is reduced.

FIG. 11a is a spectrum diagram of a surface electromyogram signal of 48 channels before noise is removed in 22 th channel. It can be seen from the spectrogram that the signal is seriously interfered by power frequency at 50 Hz; fig. 11b is a comparison graph of the frequency spectrum before and after the ICA method removes noise, fig. 11c is a comparison graph of the frequency spectrum before and after the CCA method removes noise, and fig. 11d is a comparison graph of the frequency spectrum before and after the IVA method removes noise; through comparison, the signal subjected to noise elimination by the method has the lowest energy at the frequency of 50Hz and the best effect of removing power frequency interference, and ICA and CCA methods also contain power frequency interference of different degrees. In addition, in the vicinity of other frequencies, the curves of the signal subjected to noise elimination by the method are most similar to the curves of the original signal, which shows that the method does not lose other information of the electromyographic signal while removing power frequency interference, and the ICA method and the CCA method both lose the electromyographic information to different degrees.

In conclusion, the method can solve the problem of eliminating power frequency interference and white Gaussian noise in the high-density surface electromyographic signals, and does not lose any electromyographic information. The method is suitable for preprocessing electromyographic signals such as electromyography decomposition and electromyographic prosthetic limb control, can achieve a better noise elimination effect compared with CCA and ICA methods, and has important significance for further researching activities of a neuromuscular system.